Training CNNs with Selective Allocation of Channels

Authors: Jongheon Jeong, Jinwoo Shin

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on various image classification tasks: CIFAR-10/100 (Krizhevsky, 2009), Fashion-MNIST (Xiao et al., 2017), Tiny-Image Net3, and Image Net (Russakovsky et al., 2015) datasets. We consider a variety of CNN architectures recently proposed, including Res Net (He et al., 2016a), Dense Net (Huang et al., 2017), and Res Ne Xt (Xie et al., 2017). ... In overall, our results show that training with channelselectivity consistently improves the model efficiency, mainly demonstrated in two aspects: (a) improved accuracy and (b) model compression. We also perform an ablation study to verify the effectiveness of our main ideas.
Researcher Affiliation Collaboration 1School of Electrical Engineering, KAIST, Daejeon, South Korea 2Graduate School of AI, KAIST, Daejeon, South Korea 3AITRICS, Seoul, South Korea. Correspondence to: Jinwoo Shin <jinwoos@kaist.ac.kr>.
Pseudocode Yes Algorithm 1 Channel de-allocation (dealloc)
Open Source Code No The paper does not provide any specific link to open-source code or state that code is available in supplementary materials.
Open Datasets Yes We evaluate our method on various image classification tasks: CIFAR-10/100 (Krizhevsky, 2009), Fashion-MNIST (Xiao et al., 2017), Tiny-Image Net3, and Image Net (Russakovsky et al., 2015) datasets.
Dataset Splits Yes We evaluate our method on various image classification tasks: CIFAR-10/100 (Krizhevsky, 2009), Fashion-MNIST (Xiao et al., 2017), Tiny-Image Net3, and Image Net (Russakovsky et al., 2015) datasets.
Hardware Specification No The paper does not specify the hardware (e.g., GPU or CPU models, cloud instances) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, library versions).
Experiment Setup Yes Unless otherwise stated, we fix γ = 0.001, K = 3, and Nmax = 32 for training selective convolutional layers. In cases of Dense Net-40 and Res Net164, we do not use Nmax, i.e. Nmax = , as they handle relatively fewer channels. ... The more training details, e.g. datasets and model configurations, are given in the supplementary material.